ACCROSS: Approximate Computing aCROs the System Stack

• Applicants: Professor Dr.-Ing. Hussam Amrouch, Ph.D.; Professor Dr.-Ing. Jörg Henkel; Professor Dr. Hans-Joachim Wunderlich
• Subject Area: Computer Architecture, Embedded and Massively Parallel Systems
• Term: from 2020 to 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 428566201

Computing systems have reached a point where significant improvements in computational performance and efficiency have become very hard to achieve. The main reason are power (density) and efficiency limitations due to the discontinuation of Dennard Scaling. as well as increased reliability concerns.Approximate Computing trades off precision against power, energy, storage, bandwidth or performance, and can be applied to hardware, software and algorithms. It promises to re-gain efficient computing by providing additional, adjustable design and runtime parameters to find pareto-optimal solutions. However, its application domain is still rather limited and a significant extension of the scope of applications is required, including applications that are not necessarily inherently error-tolerant.The ACCROSS project targets to tackle this challenge with a cross-layer approach to analysis and optimization, which considers the several layers (though not all) of the system stack from the application down to the hardware. At the higher levels, ACCROSS covers the analysis of applications from different computational problem classes, which are suited to act as enablers for mainstream approximate computing. This includes the development of new methods for the analysis of approximation potentials in applications, the adaptation of existing applications to approximation and the quantification and exploitation of efficiency gains. Moreover, new methods for combining suitable approximation techniques at various system layers will be provided to maximize efficiency during runtime with respect to performance and energy. Furthermore, new error metrics and methods for lightweight runtime monitoring of accuracy will be developed to ensure the amenability of approximate computing of the targeted applications. At the lower abstraction levels, ACCROSS covers the systematic evaluation of the impact of narrowing design margins which will lead to approximate behavior and increased efficiency. New models linking hardware and software will be provided, enabling designers to accurately quantify the error and efficiency tradeoff and its impact of approximation computing across the system stack.

• DFG Programme: Research Grants